New AI Masters Three Classic Board Games — By Teaching Itself How to Play!

Artificial intelligence developer DeepMind has produced an AI program that can best any human — and any game-playing program, for that matter — in a number of classic board games. This new AI, called AlphaZero, is based on an earlier program called AlphaGo, the AI that defeated Go champion Lee Sedol in 2016, but is now proficient in the games of chess, Go, and shogi. But unlike its Go-dominating predecessor, of which required extensive programming to achieve its successes, AlphaZero wasn’t programmed in its mastery of the three classic games; instead, it taught itself.

"AlphaZero just learns completely on its own, just by playing against itself," explains Julian Schrittwieser, a software engineer at DeepMind. "And we get a completely new view of the game that is not influenced by how humans traditionally play the game."
Not only did AlphaZero teach itself how to play, the self-educated AI learned these games quickly: after only taking only 34 hours to learn the game of Go, AlphaZero went on to master chess in just 9 hours. Schrittwieser says that because of AlphaZero’s generalized nature, it should be able to learn virtually any open information game — that is, any game where all of the relevant information is available to both players, unlike card games such as poker, where players are required to keep the nature of their cards secret.

Although AlphaZero wasn’t pitted against human opponents in its trial-by-board, it was matched against earlier game-playing programs that have successfully defeated human masters that were champions of their respective games. AlphaZero also made much more efficient use of its deep neural network to decide what moves it should make, making only tens of thousands of searches per second to plan a move, as opposed to the tens of millions required by its opponents, such as Stockfish, Elmo, and an earlier version of itself, AlphaGo Zero.

Despite the seemingly dry and mechanical description of how AlphaZero operates, the actual process produces game moves much like what a human would produce, despite the internal process itself being so far removed from the pattern-recognition thought processes of a person. Writing in an article for Science, chess grandmaster Garry Kasparov wrote that he was pleased that AlphaZero displayed a "dynamic, open style" similar to his own.

"The conventional wisdom was that machines would approach perfection with endless dry maneuvering, usually leading to drawn games," Kasparov wrote. "But in my observation, AlphaZero prioritizes piece activity over material, preferring positions that to my eye looked risky and aggressive.

"Programs usually reflect priorities and prejudices of programmers, but because AlphaZero programs itself, I would say that its style reflects the truth. This superior understanding allowed it to outclass the world’s top traditional program despite calculating far fewer positions per second. It’s the embodiment of the cliché, ‘work smarter, not harder.’"

And it’s this "truth" that Kasparov mentions that may be of incredible value in the future: if future AI that is designed for complex problem-solving can avoid being subject to these "priorities and prejudices" that might inadvertently (or deliberately) infect their code, then we will have an intelligence that can offer an insight from a completely different point of view, and one not hobbled by the closed loops that our own reasoning all too often falls victim to — a truly alien mind that could very well offer us a unique perspective.

Kasparov also doesn’t lament being bested in his craft by a machine, having first been beaten by IBM’s Deep Blue computer in 1997. In a December 6 Tweet, the grandmaster said "Some think I should be sad about machines getting even better, but we don’t complain about telescopes helping us see further!"